OnionNet: a Multiple-Layer Intermolecular-Contact-Based Convolutional Neural Network for Protein-Ligand Binding Affinity Prediction

Abstract

Computational drug discovery provides an efficient tool for helping large-scale lead molecule screening. One of the major tasks of lead discovery is identifying molecules with promising binding affinities toward a target, a protein in general. The accuracies of current scoring functions that are used to predict the binding affinity are not satisfactory enough. Thus, machine learning or deep learning based methods have been developed recently to improve the scoring functions. In this study, a deep convolutional neural network model (called OnionNet) is introduced; its features are based on rotation-free element-pair-specific contacts between ligands and protein atoms, and the contacts are further grouped into different distance ranges to cover both the local and nonlocal interaction information between the ligand and the protein. The prediction power of the model is evaluated and compared with other scoring functions using the comparative assessment of scoring functions (CASF-2013) benchmark and the v2016 core set of the PDBbind database. The robustness of the model is further explored by predicting the binding affinities of the complexes generated from docking simulations instead of experimentally determined PDB structures.

Figure 1

Scatter plots of the OnionNet-predicted pK_a against the experimentally measured pK_a.

Figure 2

Performance change (ΔLoss) due to missing features: (A) missing 64 features in a specific shell around the ligand; (B) missing a set of 60 features, with the same element-type combination missed in each one of the 60 “shells”. The performance change is defined as the difference between the loss of the model with missing features and the loss of the best model. The orange bars indicate the standard deviations of the ΔLoss for five independent runs.

Figure 3

Scatter plots of the predicted pK_a against the experimentally determined pK_a for the selected complexes from the v2016 core set (A) and an alignment of a re-docked native-like pose with its native pose (B). The carbon atoms in the native and native-like “good” poses for the ligand are in orange and green, respectively, whereas the oxygen and nitrogen atoms are in red and blue.

Figure 4

Featurization of the protein–ligand complexes based on contact numbers in protein–ligand interaction shells. (A) Definition of the “shell-like” partitioning of the protein around the ligand in the three-dimensional space; PDB ID 1A28 is used as an example here. (B) Glimpse of the features of the contact numbers. The features are presented column-wise, whereas the samples are presented row-wise; each row is the information we extracted from one protein–ligand complex, and one column contains a specific feature calculated from all samples.

Figure 5

Datasets used in the model. The original PDBbind v.2016 dataset was filtered to keep only the protein–ligand complexes, with measured K_i or K_d binding affinities. The remaining filtered dataset was thus divided into three disjoint datasets for training, testing, and validation. However, two overlapping testing sets were used to compare the performance of our model with other scoring functions. The numbers of protein–ligand complexes are labeled in each set in the figure.

Figure 6

Workflow of the protein–ligand binding affinity prediction with the OnionNet model.